SL.hdps.generator: SL.hdps.generator
In lendle/hdps: High-dimensional propensity score algorithm

Description Usage Arguments Details Value Author(s)

Generates a wrapper for SuperLearner using HDPS

1
2
3

SL.hdps.generator(out_name, dimension_names, predef_covar_names = c(),
  keep_k_total, ..., cvglmnet = FALSE, glmnet_args = if (cvglmnet) list()
  else list(lambda = 0))

`out_name`	Name of the outcome variable.
`dimension_names`	Dimension names of HDPS dimensions. See `hdps_screen`.
`predef_covar_names`	Names of predefined covariates to be included in logistic regression model.
`keep_k_total`	See `hdps_screen`.
`...`	Other arguments passed to `hdps_screen`.
`cvglmnet`	Use `glmnet` or `cv.glmnet` for fitting. Defaults to FALSE.
`glmnet_args`	list of arguments to be passed to glmnet or cv.glmnet. If `cvglmnet=FALSE`, `glmnet_args` should be set such that calling `predict` on the `glmnet` object returns only one vector of predictions. E.g. only one value of `lambda` should be set.

A HDPS candidate will generate covariates using hdps_screen from codes, and estimate the propensity score with logistic regression on generated covariates and predefined covariates.

To use HDPS in SuperLearner to estimate a propensity score, you need to include the outcome variable as a covariate where here outcome means the outcome of interest in the causal problem as opposed to the Y variable in SuperLearner. For non-HDPS candidates in SuperLearner, it's important to exclude the outcome variable via screen.named or some other screening algorithm in order to avoid adjusting for something downstream on the causal pathway.

A SuperLearner wrapper function

Sam Lendle

lendle/hdps documentation built on May 9, 2019, 8:34 a.m.